Optimization-AI / LibAUC

LibAUC: A Deep Learning Library for X-Risk Optimization
https://libauc.org/
MIT License
273 stars 37 forks source link

初次了解到LibAUC,希望向您请教几个问题 #38

Closed StefanIsSmart closed 9 months ago

StefanIsSmart commented 9 months ago
  1. 是否可以上线一个conda的安装途径呢?
  2. 我是否可以只使用您的Loss函数?采样、优化器都用我之前的设置,会不会对效果影响很大?(不确定的点在:a.我不能保证每个batch都能同时有正负样本,不知道这样是否是可以的;b.不知道是不是必须使用您的优化器对于计算的loss优化才算合适)。
  3. 如果必须使用您的采样器,看起来似乎不能和pyg的数据加载方式兼容是吗?

事实上我只是感觉过去的Cross Entropy不够直接优化AUC,所以想要使用一下其他的loss函数,所以最好还是能直接用您写的可微分的函数。不知道您有什么建议。

optmai commented 9 months ago
  1. Currently, we only support pip install. We may consider conda in the future.
  2. It depends on what loss do you use. If you use AUCM loss, you need to use our optimizer and data sampler. If you use other losses, e.g., partial AUC loss, AP loss, etc., you may use other optimizers. However, we have optimized our pipeline including data sampler, which may improve the performance.
  3. It can be easily adapted to other data format. You can refer to this github https://github.com/yzhuoning/DeepAUC_OGB_Challenge.
s-rog commented 2 months ago

+1 to conda install